Debate Helps Weak Judges Reward Stronger Models

In three linesDebate between models improves weak judge oversight: critic must exceed judge's classification ability for debate to help. On 5 pairings tested on code/logic tasks, 3 show statistically significant gains. Single critique suffices; rebuttal rounds add nothing. Pre-deployment audit proposed.

Summary generated by Claude — human-verified